release initial code

Co-authored-by: Ying Sheng <sqy1415@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
Co-authored-by: Zhiqiang Xie <xiezhq@stanford.edu>
Co-authored-by: parasol-aser <3848358+parasol-aser@users.noreply.github.com>
Co-authored-by: LiviaSun <33578456+ChuyueSun@users.noreply.github.com>
Co-authored-by: Cody Yu <hao.yu.cody@gmail.com>

benchmark/long_json_decode/README.md

@@ -0,0 +1,33 @@
+## Run benchmark
+
+### Benchmark sglang
+```
+python3 -m sglang.launch_server --model-path codellama/CodeLlama-7b-instruct-hf --port 30000
+```
+
+```
+python3 bench_sglang.py --num-questions 5 --parallel 1
+```
+
+
+### Benchmark vllm
+```
+python3 -m vllm.entrypoints.api_server --tokenizer-mode auto --model codellama/CodeLlama-7b-instruct-hf  --disable-log-requests --port 21000 --gpu 0.97
+```
+
+```
+python3 bench_other.py --backend vllm --num-questions 5
+```
+
+
+### Benchmark guidance
+```
+python3 bench_other.py --backend guidance --num-questions 5 --parallel 1
+```
+
+
+### Build dataset
+```
+pip install wikipedia
+python3 build_dataset.py
+```

benchmark/long_json_decode/bench_other.py

@@ -0,0 +1,104 @@
+import argparse
+import asyncio
+from concurrent.futures import ThreadPoolExecutor
+from functools import partial
+import json
+import time
+
+from tqdm import tqdm
+import numpy as np
+from sglang.test.test_utils import add_common_other_args_and_parse, call_generate_lightllm, call_generate_vllm, call_generate_srt_raw
+from sglang.utils import read_jsonl, dump_state_text
+
+
+def json_decode(document, generate):
+    s = "Please extract the information of a city from the following wikipedia page.\n"
+    s += "Page begin.\n" + document + "Page end.\n"
+    s += "Here is the name, country, and symbol of the city in JSON format.\n"
+    s += '{\n'
+    s += '  "name": "'
+    s += generate(s, max_tokens=8, stop='"') + '",\n'
+    s += '  "country": "'
+    s += generate(s, max_tokens=8, stop='"') + '",\n'
+    s += '  "air port code": "'
+    s += generate(s, max_tokens=8, stop='"') + '",\n'
+    s += '  "top 3 landmarks": "'
+    s += generate(s, max_tokens=24, stop='"') + '",\n'
+    s += '}\n'
+    return s
+
+
+def main(args):
+    lines = read_jsonl(args.data_path)
+    arguments = []
+    for i in range(len(lines[:args.num_questions])):
+        arguments.append({
+            "document": lines[i]["document"],
+        })
+    states = [None] * len(arguments)
+
+    # Select backend
+    if args.backend == "lightllm":
+        url = f"{args.host}:{args.port}/generate"
+        generate = partial(call_generate_lightllm, url=url, temperature=0)
+    elif args.backend == "vllm":
+        url = f"{args.host}:{args.port}/generate"
+        generate = partial(call_generate_vllm, url=url, temperature=0)
+    elif args.backend == "srt-raw":
+        url = f"{args.host}:{args.port}/generate"
+        generate = partial(call_generate_srt_raw, url=url, temperature=0)
+    elif args.backend == "guidance":
+        from guidance import models, gen
+
+        model = models.LlamaCpp("/home/ubuntu/model_weights/CodeLlama-7b-instruct-hf.gguf", n_gpu_layers=-1, n_ctx=11000)
+
+        def generate(prompt, max_tokens, stop):
+            out = model + prompt + gen(name="answer",
+                max_tokens=max_tokens, temperature=0, stop=stop)
+            return out["answer"]
+
+        # warmup
+        generate("Hello!", max_tokens=8, stop=None)
+    else:
+        raise ValueError(f"Invalid backend: {args.backend}")
+
+    # Run requests
+    def get_one_answer(i):
+        states[i] = json_decode(generate=generate, **arguments[i])
+
+    tic = time.time()
+    if args.parallel == 1:
+        for i in tqdm(range(len(arguments))):
+            get_one_answer(i)
+    else:
+        with ThreadPoolExecutor(args.parallel) as executor:
+            executor.map(get_one_answer, list(range(len(arguments))))
+    latency = time.time() - tic
+
+    # Compute accuracy
+    print(f"Latency: {latency:.3f}")
+
+    # Write results
+    dump_state_text(f"tmp_output_{args.backend}.txt", states)
+
+    with open(args.result_file, "a") as fout:
+        value = {
+            "task": "long_json_decode",
+            "backend": args.backend,
+            "num_gpus": 1,
+            "latency": round(latency, 3),
+            "num_requests": args.num_questions,
+            "other": {
+                "num_questions": args.num_questions,
+                "parallel": args.parallel,
+            }
+        }
+        fout.write(json.dumps(value) + "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--data-path", type=str, default="questions.jsonl")
+    parser.add_argument("--num-questions", type=int, default=100)
+    args = add_common_other_args_and_parse(parser)
+    main(args)

benchmark/long_json_decode/bench_sglang.py

@@ -0,0 +1,68 @@
+import argparse
+import json
+import time
+
+import numpy as np
+import sglang as sgl
+from sglang.test.test_utils import add_common_sglang_args_and_parse, select_sglang_backend
+from sglang.utils import read_jsonl, dump_state_text
+
+
+@sgl.function
+def json_decode(s, document):
+    s += "Please extract the information of a city from the following wikipedia page.\n"
+    s += "Page begin.\n" + document + "Page end.\n"
+    s += "Here is the name, country, and symbol of the city in JSON format.\n"
+    s += '{\n'
+    s += '  "name": "' + sgl.gen("name", max_tokens=8, stop='"') + '",\n'
+    s += '  "country": "' + sgl.gen("country", max_tokens=8, stop='"') + '",\n'
+    s += '  "air port code": "' + sgl.gen("air port code", max_tokens=8, stop='"') + '",\n'
+    s += '  "top 3 landmarks": "' + sgl.gen("landmarks", max_tokens=24, stop='"') + '",\n'
+    s += '}\n'
+
+
+def main(args):
+    lines = read_jsonl(args.data_path)
+    arguments = []
+    for i in range(len(lines[:args.num_questions])):
+        arguments.append({
+            "document": lines[i]["document"],
+        })
+
+    # Select backend
+    backend = select_sglang_backend(args)
+    sgl.set_default_backend(backend)
+
+    # Run requests
+    tic = time.time()
+    states = json_decode.run_batch(
+        arguments, temperature=0, num_threads=args.parallel)
+    latency = time.time() - tic
+
+    # Compute accuracy 
+    print(f"Latency: {latency:.3f}")
+
+    # Write results
+    dump_state_text(f"tmp_output_{args.backend}.txt", states)
+
+    with open(args.result_file, "a") as fout:
+        value = {
+            "task": "long_json_decode",
+            "backend": args.backend,
+            "num_gpus": 1,
+            "latency": round(latency, 3),
+            "num_requests": args.num_questions,
+            "other": {
+                "num_questions": args.num_questions,
+                "parallel": args.parallel,
+            }
+        }
+        fout.write(json.dumps(value) + "\n")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--data-path", type=str, default="questions.jsonl")
+    parser.add_argument("--num-questions", type=int, default=10)
+    args = add_common_sglang_args_and_parse(parser)
+    main(args)

benchmark/long_json_decode/build_dataset.py

@@ -0,0 +1,26 @@
+import json
+
+import transformers
+import wikipedia
+
+
+name = "meta-llama/Llama-2-7b-chat-hf"
+t = transformers.AutoTokenizer.from_pretrained(name)
+city_names = ["los angles", "london", "tokyo", "beijing", "singapore"]
+
+
+for city_name in city_names:
+    content = str(wikipedia.page(city_name).content)
+    content = content.replace("\n\n", "\n")
+
+    tokens = t.encode(content)
+
+    truncate_len = int((10000 / len(tokens)) * len(content))
+    truncate_content = content[:truncate_len]
+    truncate_tokens = t.encode(truncate_content)
+
+    # Count token
+    print(f"city_name: {city_name}, #tokens: {len(tokens)}, #truncate tokens: {len(truncate_tokens)}")
+
+    with open("questions.jsonl", "a") as fout:
+        fout.write(json.dumps({"document": truncate_content}) + "\n")